Main

(:toggle hide gekko button show="Show GEKKO (Python) Code":)

(:div id=gekko:)

(:source lang=python:)

from __future__ import division

from gekko import GEKKO

import numpy as np

import random

# intial parameters

n_iter = 150 # number of cycles

x = 37.727 # true value

# filtered bias update

alpha = 0.0951

# mhe tuning

horizon = 30

#%% Model

#Initialize model

m = GEKKO()

# Solve options

rmt = True # Remote: True or False

# For rmt=True, specify server

m.server = 'http://byu.apmonitor.com'

#time array

m.time = np.arange(50)

#Parameters

u = m.Param(value=42)

d = m.FV(value=0)

Cv = m.Param(value=1)

tau = m.Param(value=0.1)

#Variable

flow = m.CV(value=42)

#Equation

m.Equation(tau * flow.dt() == -flow + Cv * u + d)

# Options

m.options.imode = 5

m.options.ev_type = 1 #start with l1 norm

m.options.coldstart = 1

m.options.solver = 1 # APOPT solver

d.status = 1

flow.fstatus = 1

flow.wmeas = 100

flow.wmodel = 0

#flow.dcost = 0

# Initialize L1 application

m.solve(remote=rmt)

#%% Other Setup

# Create storage for results

xtrue = x * np.ones(n_iter+1)

z = x * np.ones(n_iter+1)

time = np.zeros(n_iter+1)

xb = np.empty(n_iter+1)

x1mhe = np.empty(n_iter+1)

x2mhe = np.empty(n_iter+1)

# initial estimator values

x0 = 40

xb[0] = x0

x1mhe[0] = x0

x2mhe[0] = x0

# outliers

for i in range(n_iter+1):

z[i] = x + (random.random()-0.5)*2.0

z[50] = 100

z[100] = 0

#%% L1 Application

## Cycle through measurement sequentially

for k in range(1, n_iter+1):

print( 'Cycle ' + str(k) + ' of ' + str(n_iter))

time[k] = k

# L1-norm MHE

flow.meas = z[k]

m.solve(remote=rmt)

x1mhe[k] = flow.model

print("Finished L1")

#%% L2 application

#clear L1//

m.clear_data()

# Options for L2

m.options.ev_type = 2 #start with l1 norm

m.options.coldstart = 1 #reinitialize

flow.wmodel = 10

# Initialize L2 application

m.solve(remote=rmt)

## Cycle through measurement sequentially

for k in range(1, n_iter+1):

print ('Cycle ' + str(k) + ' of ' + str(n_iter))

time[k] = k

# L2-norm MHE

flow.meas = z[k]

m.solve(remote=rmt)

x2mhe[k] = flow.model

#%% Filtered bias update

## Cycle through measurement sequentially

for k in range(1, n_iter+1):

print ('Cycle ' + str(k) + ' of ' + str(n_iter))

time[k] = k

# filtered bias update

xb[k] = alpha * z[k] + (1.0-alpha) * xb[k-1]

#%% plot results

import matplotlib.pyplot as plt

plt.figure(1)

plt.plot(time,z,'kx',linewidth=2)

plt.plot(time,xb,'g--',linewidth=3)

plt.plot(time,x2mhe,'k-',linewidth=3)

plt.plot(time,x1mhe,'r.-',linewidth=3)

plt.plot(time,xtrue,'k:',linewidth=2)

plt.legend(['Measurement','Filtered Bias Update','Sq Error MHE','l_1-Norm MHE','Actual Value'])

plt.xlabel('Time (sec)')

plt.ylabel('Flow Rate (T/hr)')

plt.axis([0, time[n_iter], 32, 45])

plt.show()

(:sourceend:)

(:divend:)

!!!! References

# Hedengren, J. D., Advanced Process Monitoring, Chapter accepted to Optimization and Analytics in the Oil and Gas Industry, Eds. Kevin C. Furman, Jin-Hwa Song, Amr El-Bakry, Springerâ€™s International Series in Operations Research and Management Science, 2014. [[http://apm.byu.edu/prism/uploads/Members/hedengren_apm2012.pdf|Article]]

'''Figure 1.''' Example of (1) outlier, (2) drift, and (3) noise'^[[http://dx.doi.org/10.1016/j.compchemeng.2014.04.013|1]]^'.
~~!!! Visualize time-varying data and predictions~~

(coming soon)

Data manipulation is important for dynamic optimization in order to set up simulations that utilize time-varying information. Several aspects of dynamic optimization involve the import, validation, filtering, manipulation, and display of large data sets. Select one of the following tutorials below on using MATLAB or Python to import, manipulate, and export data sets.

(:html:)

<iframe width="560" height="315" src="https://www.youtube.com/embed/E56egH10RJA" frameborder="0" allowfullscreen></iframe>

(:htmlend:)

(:html:)

<iframe width="560" height="315" src="https://www.youtube.com/embed/Tq6rCWPdXoQ" frameborder="0" allowfullscreen></iframe>

(:htmlend:)

!!! Visualize time-varying data and predictions

(coming soon)

## Dynamic Data Introduction

## Main.DynamicData History

Hide minor edits - Show changes to output

Added lines 42-181:

(:toggle hide gekko button show="Show GEKKO (Python) Code":)

(:div id=gekko:)

(:source lang=python:)

from __future__ import division

from gekko import GEKKO

import numpy as np

import random

# intial parameters

n_iter = 150 # number of cycles

x = 37.727 # true value

# filtered bias update

alpha = 0.0951

# mhe tuning

horizon = 30

#%% Model

#Initialize model

m = GEKKO()

# Solve options

rmt = True # Remote: True or False

# For rmt=True, specify server

m.server = 'http://byu.apmonitor.com'

#time array

m.time = np.arange(50)

#Parameters

u = m.Param(value=42)

d = m.FV(value=0)

Cv = m.Param(value=1)

tau = m.Param(value=0.1)

#Variable

flow = m.CV(value=42)

#Equation

m.Equation(tau * flow.dt() == -flow + Cv * u + d)

# Options

m.options.imode = 5

m.options.ev_type = 1 #start with l1 norm

m.options.coldstart = 1

m.options.solver = 1 # APOPT solver

d.status = 1

flow.fstatus = 1

flow.wmeas = 100

flow.wmodel = 0

#flow.dcost = 0

# Initialize L1 application

m.solve(remote=rmt)

#%% Other Setup

# Create storage for results

xtrue = x * np.ones(n_iter+1)

z = x * np.ones(n_iter+1)

time = np.zeros(n_iter+1)

xb = np.empty(n_iter+1)

x1mhe = np.empty(n_iter+1)

x2mhe = np.empty(n_iter+1)

# initial estimator values

x0 = 40

xb[0] = x0

x1mhe[0] = x0

x2mhe[0] = x0

# outliers

for i in range(n_iter+1):

z[i] = x + (random.random()-0.5)*2.0

z[50] = 100

z[100] = 0

#%% L1 Application

## Cycle through measurement sequentially

for k in range(1, n_iter+1):

print( 'Cycle ' + str(k) + ' of ' + str(n_iter))

time[k] = k

# L1-norm MHE

flow.meas = z[k]

m.solve(remote=rmt)

x1mhe[k] = flow.model

print("Finished L1")

#%% L2 application

#clear L1//

m.clear_data()

# Options for L2

m.options.ev_type = 2 #start with l1 norm

m.options.coldstart = 1 #reinitialize

flow.wmodel = 10

# Initialize L2 application

m.solve(remote=rmt)

## Cycle through measurement sequentially

for k in range(1, n_iter+1):

print ('Cycle ' + str(k) + ' of ' + str(n_iter))

time[k] = k

# L2-norm MHE

flow.meas = z[k]

m.solve(remote=rmt)

x2mhe[k] = flow.model

#%% Filtered bias update

## Cycle through measurement sequentially

for k in range(1, n_iter+1):

print ('Cycle ' + str(k) + ' of ' + str(n_iter))

time[k] = k

# filtered bias update

xb[k] = alpha * z[k] + (1.0-alpha) * xb[k-1]

#%% plot results

import matplotlib.pyplot as plt

plt.figure(1)

plt.plot(time,z,'kx',linewidth=2)

plt.plot(time,xb,'g--',linewidth=3)

plt.plot(time,x2mhe,'k-',linewidth=3)

plt.plot(time,x1mhe,'r.-',linewidth=3)

plt.plot(time,xtrue,'k:',linewidth=2)

plt.legend(['Measurement','Filtered Bias Update','Sq Error MHE','l_1-Norm MHE','Actual Value'])

plt.xlabel('Time (sec)')

plt.ylabel('Flow Rate (T/hr)')

plt.axis([0, time[n_iter], 32, 45])

plt.show()

(:sourceend:)

(:divend:)

Changed line 49 from:

# Hedengren, J. D., ~~Advanced Process Monitoring~~, ~~Chapter accepted to Optimization and Analytics in the Oil and Gas Industry, Eds. Kevin C. Furman, Jin-Hwa Song~~, ~~Amr El-Bakry, Springerâ€™s International Series in Operations Research and Management Science, 2014~~.~~ [[http:~~/~~/apm~~.~~byu~~.~~edu/prism/uploads/Members/hedengren_apm2012.~~pdf|Article]]

to:

# Hedengren, J. D., Eaton, A. N., Overview of Estimation Methods for Industrial Dynamic Systems, Special Issue on Optimization in the Oil and Gas Industry, Optimization and Engineering, Springer, 2015, DOI: 10.1007/s11081-015-9295-9. [[Attach:eaton_hedengren_OPTE_springer.pdf|Preprint]], [[http://link.springer.com/article/10.1007/s11081-015-9295-9|Article]]

Changed line 39 from:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as such as moving horizon estimation. There is no need to design the estimators for this problem. The estimator scripts are below with sections that can be added to simulate the effect of bad data'^1^'. Only an outlier has been added to these ~~codes~~. The ~~codes~~ should be modified to include other common phenomena such as measurement drift (gradual ramp away from the true value) and an increase in noise (random fluctuations).

to:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as such as moving horizon estimation. There is no need to design the estimators for this problem. The estimator scripts are below with sections that can be added to simulate the effect of bad data'^1^'. Only an outlier has been added to these code. The code should be modified to include other common phenomena such as measurement drift (gradual ramp away from the true value) and an increase in noise (random fluctuations). Comment on the effect of corrupted data on real-time estimation and why some methods are more effective at rejecting bad data.

Changed line 39 from:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as such as moving horizon estimation. There is no need to design the estimators for this problem. The estimator scripts are below with sections that can be added to simulate the effect of bad data'^1^'.

to:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as such as moving horizon estimation. There is no need to design the estimators for this problem. The estimator scripts are below with sections that can be added to simulate the effect of bad data'^1^'. Only an outlier has been added to these codes. The codes should be modified to include other common phenomena such as measurement drift (gradual ramp away from the true value) and an increase in noise (random fluctuations).

Added lines 41-42:

Attach:download.png [[Attach:bad_data_exercise.zip|Estimation with Outliers in MATLAB and Python]]

Changed line 45 from:

Attach:~~download~~.~~png [[Attach:bad_data_exercise.zip|Estimation with Outliers in MATLAB and Python]]~~

to:

Attach:bad_data_estimation.png

Changed lines 39-40 from:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as such as moving horizon estimation. There is no need to design the estimators for this problem. The estimator scripts are below with sections that can be added to simulate the effect of bad data.

to:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as such as moving horizon estimation. There is no need to design the estimators for this problem. The estimator scripts are below with sections that can be added to simulate the effect of bad data'^1^'.

Added lines 44-47:

!!!! References

# Hedengren, J. D., Advanced Process Monitoring, Chapter accepted to Optimization and Analytics in the Oil and Gas Industry, Eds. Kevin C. Furman, Jin-Hwa Song, Amr El-Bakry, Springerâ€™s International Series in Operations Research and Management Science, 2014. [[http://apm.byu.edu/prism/uploads/Members/hedengren_apm2012.pdf|Article]]

Changed line 39 from:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as such as moving horizon estimation.

to:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as such as moving horizon estimation. There is no need to design the estimators for this problem. The estimator scripts are below with sections that can be added to simulate the effect of bad data.

Changed lines 39-40 from:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as ~~bias updating, Kalman filter, and moving horizon ~~estimation~~. Of particular interest is the difference between the inflow (''F'_1_''') and outflow (''F'_2_''')~~.

to:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as such as moving horizon estimation.

Changed line 43 from:

Attach:download.png [[Attach:bad_data_exercise.zip|Estimation with ~~Bad Data ~~in MATLAB]]

to:

Attach:download.png [[Attach:bad_data_exercise.zip|Estimation with Outliers in MATLAB and Python]]

Changed lines 33-38 from:

The flowrate of mud and cuttings is especially important with managed pressure drilling (MPD) in order to detect gas influx or fluid losses. There are a range of measurement instruments for flow such as a [[http://en.wikipedia.org/wiki/Mass_flow_meter|mass flow meter or Coriolis flow meter (most accurate)]] and a [[http://en.wikipedia.org/wiki/Flow_measurement#Paddle_wheel_meter|paddle wheel (least accurate)]]. This particular ~~valve~~ has dynamics that are described by the following equation

0.1 d''~~F~~''~~/dt = -''F'' + C~~'~~_v_~~'~~ ~~''~~u'' + ''d''~~

with C'~~_v_~~'~~=~~1~~, ~~''~~u~~'~~' is the valve opening, and ~~''~~d~~'' ~~is a disturbance.~~

0.1 d

with C

to:

The flowrate of mud and cuttings is especially important with managed pressure drilling (MPD) in order to detect gas influx or fluid losses. There are a range of measurement instruments for flow such as a [[http://en.wikipedia.org/wiki/Mass_flow_meter|mass flow meter or Coriolis flow meter (most accurate)]] and a [[http://en.wikipedia.org/wiki/Flow_measurement#Paddle_wheel_meter|paddle wheel (least accurate)]]. This particular system has dynamics that are described by the following equation with C'_v_'=1, ''u'' is the valve opening, and ''d'' is a disturbance.

0.1 d''F'_1_'''/dt = -''F'_1_''' + C'_v_' ''u'' + ''d''

0.1 d''F'_1_'''/dt = -''F'_1_''' + C'_v_' ''u'' + ''d''

Changed line 39 from:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as bias updating, Kalman filter, and moving horizon estimation.

to:

Determine the effect of bad data (outliers, drift, and noise) on estimators such as bias updating, Kalman filter, and moving horizon estimation. Of particular interest is the difference between the inflow (''F'_1_''') and outflow (''F'_2_''').

Changed lines 33-34 from:

The flowrate of mud and cuttings is especially important with managed pressure drilling (MPD) in order to detect gas influx or fluid losses. There are a range of measurement instruments for flow such as a [[http://en.wikipedia.org/wiki/Mass_flow_meter|mass flow meter or Coriolis flow meter (most accurate)]] and a [[http://en.wikipedia.org/wiki/Flow_measurement#Paddle_wheel_meter|paddle wheel (least accurate)]].

to:

The flowrate of mud and cuttings is especially important with managed pressure drilling (MPD) in order to detect gas influx or fluid losses. There are a range of measurement instruments for flow such as a [[http://en.wikipedia.org/wiki/Mass_flow_meter|mass flow meter or Coriolis flow meter (most accurate)]] and a [[http://en.wikipedia.org/wiki/Flow_measurement#Paddle_wheel_meter|paddle wheel (least accurate)]]. This particular valve has dynamics that are described by the following equation

0.1 d''F''/dt = -''F'' + C'_v_' ''u'' + ''d''

with C'_v_'=1, ''u'' is the valve opening, and ''d'' is a disturbance.

0.1 d''F''/dt = -''F'' + C'_v_' ''u'' + ''d''

with C'_v_'=1, ''u'' is the valve opening, and ''d'' is a disturbance.

Changed line 45 from:

Attach:download.png [[Attach:bad_data_exercise.zip|Estimation with Bad Data in MATLAB~~ and Python~~]]

to:

Attach:download.png [[Attach:bad_data_exercise.zip|Estimation with Bad Data in MATLAB]]

Changed lines 27-41 from:

Some estimators and controllers are designed with ideal measurements in simulation but then fail to perform in practice due to the issues with real measurements. It is important to use methods that perform well in a variety of situations and can either reject or minimize the effect of bad data.

to:

Some estimators and controllers are designed with ideal measurements in simulation but then fail to perform in practice due to the issues with real measurements. It is important to use methods that perform well in a variety of situations and can either reject or minimize the effect of bad data.

!!!! Exercise

'''Objective:''' Understand the effect of bad data on dynamic optimization algorithms including estimator and control performance. Create a MATLAB or Python script to simulate and display the results. ''Estimated Time: 2 hours''

The flowrate of mud and cuttings is especially important with managed pressure drilling (MPD) in order to detect gas influx or fluid losses. There are a range of measurement instruments for flow such as a [[http://en.wikipedia.org/wiki/Mass_flow_meter|mass flow meter or Coriolis flow meter (most accurate)]] and a [[http://en.wikipedia.org/wiki/Flow_measurement#Paddle_wheel_meter|paddle wheel (least accurate)]].

Attach:drilling_flowrate.png

Determine the effect of bad data (outliers, drift, and noise) on estimators such as bias updating, Kalman filter, and moving horizon estimation.

!!!! Solution

Attach:download.png [[Attach:bad_data_exercise.zip|Estimation with Bad Data in MATLAB and Python]]

!!!! Exercise

'''Objective:''' Understand the effect of bad data on dynamic optimization algorithms including estimator and control performance. Create a MATLAB or Python script to simulate and display the results. ''Estimated Time: 2 hours''

The flowrate of mud and cuttings is especially important with managed pressure drilling (MPD) in order to detect gas influx or fluid losses. There are a range of measurement instruments for flow such as a [[http://en.wikipedia.org/wiki/Mass_flow_meter|mass flow meter or Coriolis flow meter (most accurate)]] and a [[http://en.wikipedia.org/wiki/Flow_measurement#Paddle_wheel_meter|paddle wheel (least accurate)]].

Attach:drilling_flowrate.png

Determine the effect of bad data (outliers, drift, and noise) on estimators such as bias updating, Kalman filter, and moving horizon estimation.

!!!! Solution

Attach:download.png [[Attach:bad_data_exercise.zip|Estimation with Bad Data in MATLAB and Python]]

Added lines 20-21:

'''Figure 1.''' Example of (1) outlier, (2) drift, and (3) noise'^[[http://dx.doi.org/10.1016/j.compchemeng.2014.04.013|1]]^'.

Added lines 15-21:

Real-data sources have a number of issues that can make simulation challenging. Measurements are used as inputs to a model, for parameter estimation, or in empirical model regression. Bad measurements can greatly affect the resulting model predictions, especially if strategies are not employed to minimize the effect of bad data.

A first step in data validation is gross error detection or when the data is clearly bad based on statistics or heuristics. Methods to automatically detect bad data include upper and lower validity limits and change validity limits. An example of a lower validity limit may be a requirement for positive (>0) values from a flow meter. Also, flow meters may not be able to detect flows above a certain limit, leading to an upper limit as well. An example of a change validity limit may be to detect sudden jumps in a measurement that are not realistic. For example, a gas chromatograph may suddenly report a jump in a gas concentration. If the gas chromatograph is measuring the concentration of a large gas phase polyethylene reactor, it is unrealistic for that concentration to change more than a certain rate. A change validity limit is able to catch these sudden shifts with gross error detection.

Other examples of real-data issues include outliers (infrequent data points that are temporarily outside of an otherwise consistent trend in the data), noise (random variations in the data due to resolution or variations in the measurement or transmission of the data), and drift (inaccurate and gradual increase or decrease of the measurement value that does not represent the true values). Data may also be infrequent (such as measurements that occur every few hours or not at regular intervals), intermittent (such as from unreliable measurements that report good values for only certain periods of time), or time delayed (measurements that are reported after a waiting period). Synchronization of real data to process models can be challenging for all of these reasons.

Some estimators and controllers are designed with ideal measurements in simulation but then fail to perform in practice due to the issues with real measurements. It is important to use methods that perform well in a variety of situations and can either reject or minimize the effect of bad data.

A first step in data validation is gross error detection or when the data is clearly bad based on statistics or heuristics. Methods to automatically detect bad data include upper and lower validity limits and change validity limits. An example of a lower validity limit may be a requirement for positive (>0) values from a flow meter. Also, flow meters may not be able to detect flows above a certain limit, leading to an upper limit as well. An example of a change validity limit may be to detect sudden jumps in a measurement that are not realistic. For example, a gas chromatograph may suddenly report a jump in a gas concentration. If the gas chromatograph is measuring the concentration of a large gas phase polyethylene reactor, it is unrealistic for that concentration to change more than a certain rate. A change validity limit is able to catch these sudden shifts with gross error detection.

Other examples of real-data issues include outliers (infrequent data points that are temporarily outside of an otherwise consistent trend in the data), noise (random variations in the data due to resolution or variations in the measurement or transmission of the data), and drift (inaccurate and gradual increase or decrease of the measurement value that does not represent the true values). Data may also be infrequent (such as measurements that occur every few hours or not at regular intervals), intermittent (such as from unreliable measurements that report good values for only certain periods of time), or time delayed (measurements that are reported after a waiting period). Synchronization of real data to process models can be challenging for all of these reasons.

Some estimators and controllers are designed with ideal measurements in simulation but then fail to perform in practice due to the issues with real measurements. It is important to use methods that perform well in a variety of situations and can either reject or minimize the effect of bad data.

Deleted lines 14-16:

(coming soon)

Added lines 4-17:

Data manipulation is important for dynamic optimization in order to set up simulations that utilize time-varying information. Several aspects of dynamic optimization involve the import, validation, filtering, manipulation, and display of large data sets. Select one of the following tutorials below on using MATLAB or Python to import, manipulate, and export data sets.

(:html:)

<iframe width="560" height="315" src="https://www.youtube.com/embed/E56egH10RJA" frameborder="0" allowfullscreen></iframe>

(:htmlend:)

(:html:)

<iframe width="560" height="315" src="https://www.youtube.com/embed/Tq6rCWPdXoQ" frameborder="0" allowfullscreen></iframe>

(:htmlend:)

!!! Visualize time-varying data and predictions

(coming soon)

Added lines 1-3:

(:title Dynamic Data Introduction:)

(:keywords dynamic data, validation, simulation, modeling language, differential, algebraic, tutorial:)

(:description Dynamic data for Differential Algebraic Equation (DAE) systems for use in dynamic simulation, estimation, and control:)

(:keywords dynamic data, validation, simulation, modeling language, differential, algebraic, tutorial:)

(:description Dynamic data for Differential Algebraic Equation (DAE) systems for use in dynamic simulation, estimation, and control:)